UvA-MT improves machine translation by refining and judging outputs
UvA-MT's Participation in the WMT26 General Translation Shared Task
Computation and Language
Summary
Machine translation aims to turn sentences from one language into another accurately. UvA-MT's team improved this by first generating many translation options for each sentence using a step-by-step approach. Then, they used large language models to compare these options and pick the best one through voting rounds. Their tests showed this method works better than simply picking the best single option directly.
What this means in practice
- •For machine translation developers: Improve translation quality by generating and refining multiple candidate translations and selecting the best through model-based pairwise comparison.
- •For multilingual content managers: Enhance automated translation workflows across many language pairs by incorporating refinement and voting techniques to increase accuracy.
Authors
Di Wu, Sergey Troshin, Wafaa Mohammed, Seth Aycock, Evgeniia Tokarchuk, Vlad Niculae, Christof Monz
Abstract
This paper presents UvA-MT's submission to the WMT 2026 General Machine Translation shared task, competing in the unconstrained track across all 23 translation directions. This year, we fully leverage the test-time methods of Large Language Models (LLMs) for machine translation. Specifically, (1) we use sequential sampling with a refinement prompt to generate a pool of translation candidates for each source segment, a strategy shown to outperform traditional parallel sampling; and (2) we employ LLMs as pairwise judges to select the best candidate via a round-robin voting mechanism, with offline evaluation on the WMT25 benchmark showing a very clear improvement over point-wise best-of-N selection. Combined, these two strategies form our WMT26 submission.